ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-04-26 13:59:43 +00:00
Code Issues Actions 4 Packages Projects Releases Wiki Activity

llama : fix buffer checks for mamba and rwk (llama/10111)

Browse Source 
* llama : fix buffer checks for mamba and rwk

* llama : fix missing worst case flag during reserve

* cuda : fix supports_op for norm

* disable sched SET_CAUSE
This commit is contained in:
Diego Devesa 2024-10-31 22:54:23 +01:00 committed by Georgi Gerganov
parent 371bfaca8c
commit 3e231ab9cc
3 changed files with 6 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-backend.cpp
View File

@ -1508,7 +1508,7 @@ static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, co
return -1; return -1;
} }
#if 1 #if 0
#define GGML_SCHED_MAX_SPLITS_DEBUG 4096 #define GGML_SCHED_MAX_SPLITS_DEBUG 4096
static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS_DEBUG*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS_DEBUG*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only
#define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__) #define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__)

6
ggml/src/ggml-cuda.cu
View File

@ -3107,18 +3107,20 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
} }
return false; return false;
} break; } break;
case GGML_OP_NORM:
case GGML_OP_RMS_NORM:
return ggml_is_contiguous(op->src[0]) && op->ne[0] % WARP_SIZE == 0;
break;
case GGML_OP_NONE: case GGML_OP_NONE:
case GGML_OP_RESHAPE: case GGML_OP_RESHAPE:
case GGML_OP_VIEW: case GGML_OP_VIEW:
case GGML_OP_PERMUTE: case GGML_OP_PERMUTE:
case GGML_OP_TRANSPOSE: case GGML_OP_TRANSPOSE:
case GGML_OP_NORM:
case GGML_OP_ADD: case GGML_OP_ADD:
case GGML_OP_ADD1: case GGML_OP_ADD1:
case GGML_OP_SUB: case GGML_OP_SUB:
case GGML_OP_MUL: case GGML_OP_MUL:
case GGML_OP_DIV: case GGML_OP_DIV:
case GGML_OP_RMS_NORM:
case GGML_OP_SCALE: case GGML_OP_SCALE:
case GGML_OP_SQR: case GGML_OP_SQR:
case GGML_OP_SQRT: case GGML_OP_SQRT:

1
ggml/src/ggml.c
View File

@ -7243,6 +7243,7 @@ struct ggml_tensor * ggml_ssm_conv(
const int64_t n_s = sx->ne[2]; const int64_t n_s = sx->ne[2];
// TODO: maybe support other strides than 1? // TODO: maybe support other strides than 1?
// FIXME: this is always true?
GGML_ASSERT(sx->ne[0] == d_conv - 1 + n_t); GGML_ASSERT(sx->ne[0] == d_conv - 1 + n_t);
GGML_ASSERT(sx->ne[1] == d_inner); GGML_ASSERT(sx->ne[1] == d_inner);
GGML_ASSERT(n_t >= 0); GGML_ASSERT(n_t >= 0);